Why it matters
FLUX took over the role Stable Diffusion had played in the open image ecosystem. Community tools, fine-tunes and LoRAs moved to it, and it's widely available in ComfyUI, on inference platforms such as Replicate and fal, and inside commercial products. Its API models are also used behind the scenes by a number of well-known creative apps.
It also pushed the field toward a new architecture. FLUX uses a diffusion transformer trained with flow matching (specifically rectified flow), rather than the U-Net diffusion design of earlier Stable Diffusion versions. That combination has since become the standard recipe for top image and video models.
Deep Dive
FLUX is made by Black Forest Labs, a German lab founded by several of the researchers behind Stable Diffusion. The first family, FLUX.1, arrived in August 2024 in three versions: Pro (via API only), Dev (open weights for non-commercial use) and Schnell ("fast" in German, a distilled version released under the permissive Apache 2.0 license).
FLUX quickly became the leading open-weight image model. It followed long, detailed prompts better than earlier open models, rendered hands and text far more reliably, and matched or beat many closed models of its time.
The Model Family
- FLUX.1 [pro], [dev], [schnell] (2024): the original text-to-image models, about 12 billion parameters. Schnell generates good images in just 1 to 4 steps.
- FLUX.1 Tools: add-ons for inpainting, outpainting and structure control (depth or edge maps), similar in spirit to ControlNet.
- FLUX.1 Kontext (2025): an image editing model. You give it an image plus an instruction ("change the jacket to red", "put this character on a beach") and it edits while keeping the subject consistent across edits.
- FLUX.2 (late 2025): the next generation, with higher resolution, better text, multi-reference image input and again a mix of API and open-weight versions.
Under the Hood
FLUX.1 generates in the compressed latent space of a 16-channel VAE, which preserves more detail than the 4-channel latents of older Stable Diffusion models, and FLUX.2 uses a 32-channel VAE. Prompts are read by large text encoders (T5 alongside CLIP in FLUX.1, a single Mistral text encoder in FLUX.2), which is a big reason it understands complex prompts. The Dev and Schnell versions were distilled from Pro, so they don't use classic negative prompts or classifier-free guidance in the usual way; instead they take a "guidance" value built into the model.
Running It Yourself
FLUX.1 Dev at full precision needs about 24 GB of VRAM, but quantized versions (FP8, NF4 or GGUF) run on 8 to 12 GB cards with some quality trade-off. ComfyUI is the most common way to run it locally, and fine-tuning LoRAs for a specific style, product or person is a popular use.
Licensing
Check the license for each version. Schnell is Apache 2.0 and usable commercially. Dev-tier weights use a non-commercial license, with commercial use available through Black Forest Labs' API or a paid license.